Parallelization of a meteorological model using message passing interface and CUDA: A case study with the inversion estimation algorithm

Abstract

AbstractSome algorithms deal with large amounts of data and vast and complex calculations, such as the algorithms used in numerical weather models in meteorological science. The need for extensive computer and computing resources is a problem in the process of running meteorological models, and this problem cannot be solved by or is not efficient to run on classical computing. Instead of traditional computers, high‐performance computer systems are used as a solution to this problem with the Message Passing Interface (MPI) to split the computational load over many CPU cores. In this study, we undertake the parallelizing of a meteorological algorithm written for the special purpose of forecasting temperature inversion. We evaluate the performance of running this algorithm in parallel to see the effect of using multiple cores on the performance with a specific example. Furthermore, we evaluate the algorithm's memory consumption by running the program on a GPU with CUDA and measure the execution times. The results show that the execution time decreases when using more cores until a breaking point, and this breaking point is changed proportionally with the size of the algorithm. In terms of time, we obtain up to 84.95% better performance by running the program in parallel using MPI and up to 94.10% better performance by running the program in parallel using CUDA.